CONTINUAL INTELLIGENCEAn illustrated essay series
About

A field guide to learning that never stops.

Continual Intelligence is a twelve-part illustrated essay series. It reads the primary literature closely — and tries to leave you with a picture, not just a citation.

The premise

Most machine learning assumes a world that holds still long enough to be measured: a fixed dataset, a training phase, an evaluation phase, a number. The systems we actually want — agents in deployment, models that improve after release, reasoners that face problems no one curated — live in a world that never stops moving.

This series takes that seriously. If the world is bigger than the agent, then learning cannot have a completion point, forgetting is sometimes correct, and the benchmarks we trust may be measuring the wrong thing entirely. Each essay follows one such thread from its primary sources to a claim you can argue with.

How to read it

The essays are numbered and build on one another, but each stands on its own. Start at the benchmark gap if you want the argument in order, or jump to whatever pulls you. Every essay carries its anchor papers up top, key takeaways inline, and full references at the end — arXiv links included.

The research threads

Seven tags trace the threads that weave through the series:

  • CL — Continual Learning
  • RL — Reinforcement Learning
  • WM — World Models
  • RE — Reasoning
  • GI — General Intelligence
  • EV — Evolutionary & Open-Ended
  • FN — Foundations

On the illustrations

The pictures are deliberately pixel-art: blocky, low-resolution, honest about being diagrams rather than photographs. A continual learner facing a vast grid world; a network fracturing as it loses plasticity; a chain of thought folding into a spiral. The aim is the same as the prose — a clear silhouette of an idea.

The full index

  1. The Benchmark Gap in Continual RL: From Continual World to SPIRAL
  2. The Plasticity Crisis in Continual Deep Learning
  3. The Big World Hypothesis: Why Continual Learning Is Inevitable
  4. GVFs as Proto-World-Models: The Alberta Plan Vindicated?
  5. The Forgetting Transformer: When Architecture Solves Plasticity
  6. Does RL Teach LLMs to Reason, or Just Refine Them?
  7. Shape of Thought: Why Reasoning Format Matters More Than Correctness
  8. Stable Deep RL at Scale: Gradients, KL, and the Shape of Learning
  9. Reasoning at Scale: What DeepSeek-R1, ProRL, and Prolonged RL Reveal
  10. Darwin-Gödel to ShinkaEvolve: The Case for Open-Ended AI
  11. Thinking Without Tokens: CTM and Inference-Time Compute Beyond CoT
  12. RL as Educator: Training Teachers, Not Just Students